Introduction To Simulation

Proceedings of the 2013 Winter Simulation Conference
R. Pasupathy, S.-H. Kim, A. Tolk, R. Hill, and M. E. Kuhl, eds
INTRODUCTION TO SIMULATION
Ricki G. Ingalls
Oklahoma State University

322 Engineering North
Stillwater, OK 74078, USA
ABSTRACT
Simulation is a powerful tool if understood and used properly. This introduction to simulation tutorial is
designed to teach the basics of simulation, including structure, function, data generated, and its proper
use. The introduction starts with a definition of simulation, goes through a talk about what makes up a
simulation, how the simulation actually works, and how to handle data generated by the simulation.
Throughout the paper, there is discussion on issues concerning the use of simulation in industry.
1 DEFINITION OF SIMULATION
Simulation, according to Shannon (1975), is the process of designing a model of a real system and con-
ducting experiments with this model for the purpose either of understanding the behavior of the system or
of evaluating various strategies (within the limits imposed by a criterion or set of criteria) for the opera-
tion of the system. According to this definition, a simulation can be a discrete-event simulation, as we
will discuss in this paper. Many people who attend this conference will be familiar with the term MRP
simulation. This is a model (actually a copy) of the real system (the MRP system of record) on which
experiments (or scenarios) can be run to evaluate various strategies (such as how to respond to a drastic
change in the forecast). Although we do not teach courses in our curriculum on MRP simulation, a
MRP simulation is no less a simulation than the type of simulation we will discuss in this tutorial.
The difference, and the power, of discrete-event simulation is the ability to mimic the dynamics of a
real system. Many models, including high-powered optimization models, cannot take into account the
dynamics of a real system. It is the ability to mimic the dynamics of the real system that gives discrete-
event simulation its structure, its function, and its unique way to analyze results. So, to take liberties with
one of my mentors in this field, we will say that simulation is the process of designing a dynamic model
of an actual dynamic system for the purpose either of understanding the behavior of the system or of eval-
uating various strategies (within the limits imposed by a criterion or set of criteria) for the operation of the
system.
Throughout this paper, we will be referencing ideas and thoughts from Shannon (1975), Law (2007),
Banks, et. al. (2009), Kelton, Sadowski, and Sadowski (2001), Ingalls and Kasales (1999), Ingalls (1998),
and Ingalls and Eckersley (1992).
2 A DRIVE-THROUGH EXAMPLE
In order to give us a reference model for discussion purposes, let us consider the example of a drive-
through window at a fast food restaurant whose logic is shown in Figure 1. As a car enters from the
street, the driver, who we will call Fred, decides whether or not to get in line. If Fred decides to leave the
restaurant, he leaves as a dissatisfied customer. One of the great things about a simulation is that it is easy
978-1-4799-2076-1/13/$31.00 2013 IEEE 291

Ingalls
Enter Restaurant
Parking Lot
Leave the
Is the Line Too Resturant as a
TRUE
Long? dissatisfied
customer
FALSE
Get in the Line for

the Menu Board
When Arrive at the

Order is Made in
Menu Board, Give ORDER
the Kitchen
Order
CAR
Is there Room Wait until the Car

to Move FALSE In Front of you
Forward? Moves
TRUE
Get in Line for

Order Pickup
Window Order Meets
Car at Order
Pickup Window
Arrive at Order
Pickup Window -
Pay and Get Order
Leave the
Restaurant Parking
Lot
Figure 1: Drive-Through Example Logic Flow
to track this type of customer. In most real-world systems, it is usually difficult to track customers who
leave dissatisfied.
If Fred decides to get in line, then he waits until the menu board (with the speaker no normal human
being can understand) is available. At that time, Fred gives the order to the order taker.
292
Ingalls
After the order is taken, then two things occur simultaneously. (1) Fred moves forward if there is
room. If there is no room, then he has to wait at the menu board until there is room to move forward. As
soon as there is room, he moves so the next customer can order. (2) The order is sent electronically back
to the kitchen where it is prepared as soon as the cook is available.
As soon as Fred reaches the pickup window, then he pays and picks up his food, if it is ready. If the
food is not ready, then Fred has to wait until his order is prepared. In this drive-through, Fred never
hears, Please pull forward and we will bring your order out to you. That is one of the reasons that Fred
and I both like this place.
As soon as Fred parts with his money and gets his food, then he leaves the pickup window a satisfied
customer.
3 DISCRETE-EVENT SIMULATION STRUCTURE
Although there are various flavors and paradigms in discrete-event simulation, there has evolved a basic
structure that is used by most simulation packages. Regardless of how complex a discrete-event simula-
tion package may be, it is likely to contain the basic components that we will describe in this section.
The structural components of a discrete-event simulation include entities, activities and events, re-
sources, global variables, a random number generator, a calendar, system state variables and statistics
collectors.
3.1 Entities
The best way to understand the function of an entity is understand that entities cause changes in the state
of the simulation. Without entities, nothing would happen in a simulation. As a matter of fact, one stop-
ping condition for a simulation model is the condition where there are no active entities in the system.
Entities have attributes. Attributes are characteristics of a given entity that are unique to that entity.
Attributes are critical to the understanding of the performance and function of entities in the simulation.
In our example, the primary entity type is the cars coming to the restaurant. Also, there is a second
entity type created when the order is taken, and that is the order itself. It has a relatively short life in the
simulation, lasting only from the time the order is taken until it meets back up with the car at the pickup
window.
We also have two attributes in our simulation. The first one is the time of day that the car enters the
restaurant parking lot. We will call this attribute StartTime. The second is the value of the order. We
will call this attribute OrderValue. Both of these are unique to the customers who are represented in the
simulation.
Another example of an entity is the part that is flowing through a factory. Entities that represent parts
in a factory could be created randomly or according to a schedule. A common attribute would be the time
that the part started in the factory. Each part would have a unique time that it started in the factory. It
may also have other attributes such as priority, the type of part, and cost incurred to produce the part. In a
normal factory simulation that is tracking each individual part, it would not be unusual to have thousands
of entities active in the simulation simultaneously.
Entities, however, do not need to be parts in a manufacturing facility or cars in a drive-through. It is
also a common practice to use entities to represent the flow of information. This information could be a
customer order, an email alert, a packet in a computer network, etc. It can be anything non-physical that
causes a change in the status of system. These entities also have attributes. For example, an entity repre-
senting a packet in a computer network would have attributes such as packet size, packet destination, the
time that the packet would time-out, etc.
293
Ingalls
3.2 Activities and Events
Activities are processes and logic in the simulation. Events are conditions that occur at a point in time
which cause a change in the state of the system. An entity interacts with activities. Entities interacting
with activities create events.
There are three major types of activities in a simulation: delays, queues and logic. The delay activity
is when the entity is delayed for a definite period of time. In our example, there are three delays. The
first is when Fred is ordering at the menu board, the second is when the order is being cooked in the
kitchen, and the third is Fred picks up his order at the pickup window. In general, the length of time for a
delay is either constant or is randomly generated. At the point that the entity starts the delay, an event
occurs. This event schedules the entity on the calendar (which we will get to later). If the delay is for d
time units, then the entity is scheduled to complete the delay d time units after the current time of the
simulation. At that time, the delay expires and another event is generated.
Queues are places in the simulation were entities wait for an unspecified period of time. Entities can
be waiting on resources (which we will get to later) to be available or for a given system condition to oc-
cur. Queues are most commonly used for waiting in line for a resource or storing material that will be
taken out of the queue when the right conditions exist. In our example, there are three queues, the first is
the part of the line that waits on the menu board to be available, the second is the order waiting on the
kitchen becoming available, and the third is the part of the car line waiting on the order pickup window to
become available. All three of these queues are waiting for resources to be available.
Logic activities simply allow the entity to effect the state of the system through the manipulation of
state variables (which we will get to later) or decision logic. The first of several logic activities in our ex-
ample is the decision whether or not to get in the order line in the first place. This decision is effected by
the length of the line in front of the menu board.
3.3 Resources
In a simulation, resources represent anything that has a restricted (or constrained) capacity. Common ex-
amples of resources include workers, machines, nodes in a communication network, traffic intersections,
etc. In our example, we have three different resources. The first resource is the menu board, which only
one car can use at a time. The menu board is utilized from the time a car moves in front of it until the car
moves away from it. In our model, the car does not automatically move away from the menu board after
the order has been placed. There must also be room to move forward. So, this resource is occupied for
both productive time (when the order is being placed) and unproductive time (when there is not enough
room to pull the car forward). The second resource is the kitchen. It is utilized from the time an order ar-
rives until the order is finished cooking. The third resource is the order pickup window. This resource
can also be unproductive. That occurs when the car has arrived at the window, but the order is not ready
from the kitchen yet. In the course of the simulation, we can track the key statistics of each of these re-
sources, including utilization and costs.
It should also be noted that very complex resources can be utilized in a simulation. In a manufactur-
ing simulation, conveyors are a very complex resource that many simulation packages offer. Also, trans-
portation options such as trucks are offered as resources. A third complex resource is a vat or container
that has a (continuous) flow of material both in an out of the resource. Depending on the target market of
the simulation package, many complex resources are available to use.
3.4 Global Variables
If you are a programmer, then the idea of having global variables is nothing new. A global variable is a
variable that is available to the entire model at all times. A global variable can track just about anything
that is of interest to the entire simulation. In our model we have four global variables, two of which help
294
Ingalls
us configure the problem and two of which collect revenue information. The two variables that help us
configure the problem are the length of the line allowed at the restaurant. The first is MenuBoard-
LineSize, which gives the maximum line size for the menu board, including the car who is at the menu
board. The second is OrderPickupLineSize, which gives the maximum line size after the menu board, in-
cluding the car at the pickup window. By changing these two variables, the simulator can analyze the ef-
fectiveness of different line configurations.
The two variables that help use collect revenue information are Revenue and LostRevenue. Revenue
is the total amount of revenue collected in the simulation, and LostRevenue is the total amount of revenue
lost because the customer thought that the line was too long.
3.5 Random Number Generator
Every simulation package has a random number generator. The random number generator (technically
called a pseudo-random number generator) is a software routine that generates a random number between
0 and 1 that is used in sampling random distributions. For example, let us assume that you have deter-
mined that a given process delay is uniformly distributed between 10 minutes and 20 minutes. Then eve-
ry time an entity went through that process, the random number generator would a generate a number be-
tween 0 and 1 and evaluate the uniform distribution formula that has a minimum of 10 and a maximum of
20. As an example, let us assume that the generated random number is 0.7312, then the delay would be
10+(0.7312)*(20-10) = 17.312. So the entity would delay for 17.312 time units in the simulation. Every-
thing that is random in the simulation uses the random number generator as an input to determine values.
In our example model, we have five randomly assigned delays and values in the model, (1) the time
between arrivals of cars to the restaurant, (2) the value of the order for the car, (3) the delay at the menu
board, (4) the delay at the kitchen, and (5) the delay at the order pickup window. All of these delays are
uniformly distributed and their ranges are as follows:
1. Time between arrivals of cars to the restaurant = 20 to 120 seconds
2. The value of the order for the car = 2 to 22 dollars.
3. The delay at the menu board = 20 to 120 seconds.
4. The delay at the kitchen = 16 to 96 seconds.
5. The delay at the order pickup window = 20 to 120 seconds.
3.6 The Calendar
The calendar for the simulation is a list of events that are scheduled to occur in the future. In every simu-
lation, there is only one calendar of future events and it is ordered by the earliest scheduled time first. In
a later example, it will become more clear how the calendar works and why it is important in the simula-
tion. At this point, just remember that, at any given point in time, every event that has already been
scheduled to occur in the future is held on the calendar.
3.7 System State Variables
Depending on the simulation package, there can be several system state variables, but the one system state
variable that every simulation package has is the current time of the simulation. In order to keep from of-
fending any simulation vendors, we will choose a different name for our simulation time variable. Our
name of the current time of the simulation is CurrentTime. This variable is updated every time an entity
is taken from the calendar.
3.8 Statistics Collectors
Statistics collectors are a part of the simulation that collects statistics on certain states (such as the state of
a resources), or the value of global variables, or certain performance statistics based on attributes of the
295
Ingalls
entity. There are three different types of statistics that are collected, counts, time-persistent, and tallies.
Counts are very straightforward, they count. In our model, we count the number of Lost Customers be-
cause the line was too long. Time-persistent statistical collectors give the time-weighted values of differ-
ent variables in the simulation. A common variable to track is the utilization of a resource. In our model,
we collect 6 different time-persistent statistics, which are the number of busy resources of the three re-
sources that we have in the model and, the number of entities in each of the three queues that serve the
three resources. Tally statistics are collected one observation at a time without regard to the amount of
time between observations. In our model, we collect a very common statistic, which is the amount of
time that an entity stays in the system. Since we assign the value of StartTime to the attribute of the entity
when it enters the restaurant parking lot, then the value CurrentTime StartTime is the total amount of
time in the system. If we are to improve this system, we would want to minimize this statistic without
hurting the revenue too much.
4 A WALK THROUGH THESE CONCEPTS
What we are going to do now is walk through a couple of steps in our simulation model. As you will see,
we will track the state of the system, the entities on the calendar, the values of attributes and state varia-
bles, and the statistics we are collecting.
4.1 The State of The System at Noon
Figure 2 shows the state of the system at noon during a lunch rush at our fast food restaurant. To help us
design the line at the restaurant, we have set two of our global variables, MenuBoardLineSize and Order-
PickupLineSize, both equal to 3. As you can see, that is the total capacity of the line at the restaurant.
The state of our system at noon is as follows:
Car 1 (the red Ferrari) is at the Order Pickup Window receiving its order.
Cars 2 (the pink Cadillac) and 3 (the blue Fiesta ) are in line waiting for the Order Pickup Window.
Car 4 (the blue Smart car) is ordering at the Menu Board.
The Order for Car 2 is waiting to be
picked up.
The Order for Car 3 is cooking in
the kitchen
Car 5 (of undetermined type) is
schedule to arrive at the restaurant
in the future.
The calendar of this system is made up
of the entities that are scheduled to com-
plete an activity with a specific time du-
ration. These entities are Car 1, Car 4,
the Order for Car 3, and Car 5. The cal- Figure 2: State of the System at Noon
endar for these four entities is shown in
Table 1.
296
Ingalls
Table 1: The Calendar at Noon
Entity Event Event Time

Car(1) Order Pickup Window Complete 12:00:38 PM
Car(5) Car Enters System 12:00:46 PM
Car(4) Menu Board Complete 12:00:49 PM
Order(2) Kitchen Order Complete 12:01:01 PM
We also know the values of the attributes of the entities in the system. As you recall, we have two at-
tributes, StartTime and OrderValue. The values for those attributes for each of the entities in the system
is outlined in Table 2.
Table 2: Entity Attributes at Noon

Entity StartTime OrderValue
Car(1) 11:55:08 AM $18.13
Car(2) 11:55:40 AM $17.90
Car(3) 11:58:00 AM $21.03
Car(4) 11:58:53 AM $13.72
Car(5) --- ---
Other important information concerns our system state variable, CurrentTime, which is set to
12:00:00 PM.
The statistics that we are tracking in the simulation have the values listed in Table 3 as of noon. The
Time/Obs column gives the amount of time that we have been collecting the statistic (which is since
11:00 AM) for time-persistent statistics or the number of observations for tally statistics.
Table 3: Statistics at Noon

Statistic Value Time/Obs
Revenue $513.52
Lost Revenue $31.94
Lost Customers 3
MenuBoard Utilization 98.88% 1:00:00
Kitchen Utilization 71.90% 1:00:00
OrderWindow Utilization 95.19% 1:00:00
MenuBoard Waiting in Line 0.842 1:00:00
Kitchen Waiting in Line 0.243 1:00:00
OrderWindow Waiting in Line 1.224 1:00:00
Time In System 5.0692 45
4.2 The State of The System at 12:00:38 PM
Here is where we get one of the key ideas about a discrete-event simulation. A discrete-event simulation
moves the current time of the simulation through events and their timing instead of distinct time intervals.
If we had distinct time intervals of 1 second, we would go through 38 time intervals before anything
would happen. So instead of going through 38 time intervals with nothing happening, we go straight to
297
Ingalls
the next scheduled event, which is the first event on the calendar, which is scheduled to occur at 12:00:38
PM. That event is Car 1 being done at the order pickup window.
When Car 1 is finished at the order pickup window, there are several activities that take place. The
first thing to do is set the CurrentTime to 12:00:40 PM. After that the Revenue is increased by the value
of Car 1s order. So, Revenue increases from $513.52 to $531.65. The second activity that occurs that
that Car 1s Time In System is recorded. Since Car 1 arrived at 11:55:08 AM, its Time In System is
12:00:38 11:55:08 = 5:30. Then Car 1 leaves the system and its entity is destroyed.
Also, when Car 1 no longer has the
Order Pickup Window resource busy, Car
2 can allocate that resource and start its
process of picking up the order. Logically,
Car 2 has to check to see if its order is al-
ready back from the kitchen, and it is. So,
Car 2 can immediately begin his delay at
the Order Pickup Window. The delay is Figure 3: The State of the System at 12:00:38 PM
randomly set to 48 seconds, so Car 2 will
finish at the Order Pickup Window at
12:01:16 PM. The resulting state of the system is shown in Figure 3.
After Car 1 leaves and Car 2 starts its delay at the Order Pickup Window, we have a new calendar,
which is shown in Table 4.
Table 4: The Calendar at 12:00:38 PM

Entity Event Event Time
Car(5) Car Enters System 12:00:46 PM
Car(4) Menu Board Complete 12:00:49 PM
Order(2) Kitchen Order Complete 12:01:01 PM
Car(2) Order Pickup Window Complete 12:01:16 PM
The statistics can be updated at this point as well. However, most simulation packages only update
statistics when the value that is being tracked changes. We have two types of statistics, the counting of
Revenue and RevenueLost, the resource utilization statistics and the queue length statistics. As we men-
tioned above, Revenue was increased from $513.92 to $531.65. The other statistics are time-dependent
statistics that are time-weighted averages of a given value. As an example, let us calculate the new value
of the average number of cars waiting in line for the menu board. At noon, the simulation had been run-
ning for one hour and the average value was 0.842. From noon to 12:00:20 PM, the number of cars wait-
ing in line for the menu board has been 0. So the new time-weighted average is ((0.842 * 1:00:00) + (0 *
0:00:38)) / 1:00:38 = 0.833. All time-dependent statistics are calculated in this way. The value for all of
the statistics at time 12:00:38 PM is in Table 5.
298
Ingalls
Table 5: Statistics at 12:00:38 PM

Revenue $531.65
Lost Revenue $31.94
Lost Customers 3
4.3 The State of The System at 12:00:46 PM
Lets take one more step in the simulation.

Looking at Table 4, we know that the next
scheduled event is that Car 5 enters the sys-
tem. The first thing to do is set the Cur-
rentTime to 12:00:40 PM. When Car 5 Ar-
rives, it is given 2 is assigned the time it
enters the system and also assigned a random
OrderValue. In this case the OrderValue is
$9.85. Figure 4: The State of the System at 12:00:46 PM
Car 5 then looks to see if it can enter the
MenuBoard queue. Since there is room, then Car 5 waits behind Car 4 in line. The state of the system at
12:00:46 is shown in Figure 4. Also, the statistics at 12:00:46 are shown in Table 6.
Table 6: Statistics at 12:00:46 PM

Revenue $531.65
Lost Revenue $31.94
Lost Customers 3
5 INTERPRETING OUTPUT STATISTICS
Let us assume that our model has run from 11:00 AM to 2:00 PM. At 2:00 PM, we stop the simulation
and we have all of our statistics calculated. The answer for the simulation is in Table 7.
299
Ingalls
Table 7: Statistics at End of Simulation

Revenue $1616.12
Lost Revenue $106.73
Lost Customers 9
But what does this data really telling us? Is it saying that every day, from 11:00 AM to 2:00 PM, the
revenue for the restaurant will be $1,616.12? Is it saying that every day, from 11:00 AM to 2:00 PM, the
average amount of time that a car will be in line is 4.8951 minutes? What is this really saying?
The answer to that question is, These numbers give us a random answer to the performance of the
system. It is a random answer because there are random inputs to the system that give us this answer.
So how can we be sure that the answer is any good?
The answer generated by only one run is not really an answer at all. To make the point, let us take the
following 100 rolls of the dice. You would think that 100 rolls of the dice would tell us the average. But
if you take the 100 rolls in Table 8, the average is only 6.72. Is that close enough? Would you be willing
to bet that the next 100 rolls would come up with an average of 6.72. Probably not. (If you are, please
contact me as soon as possible on any other wagers you may be considering.)
Table 8: 100 Rolls of the Dice

6 5 8 6 5 4 10 6 7 8
5 9 8 7 8 6 3 8 7 6
9 9 8 8 6 8 9 7 10 5
2 10 11 8 6 8 7 3 8 8
4 6 8 11 2 4 8 9 8 5
9 3 8 7 2 3 9 10 7 7
3 9 5 7 7 7 9 4 8 7
4 10 7 4 10 8 4 8 9 7
3 6 6 3 6 3 10 9 7 4
6 8 5 9 12 6 8 6 4 2
Average: 6.72
So what is the answer to this problem. The answer is that if we really want to estimate the average
value of the roll of the dice if we roll the dice 100 times, we need to run the simulation more times. Each
time that we run a simulation is called an iteration. So, as an example, let us say that we have run our
dice throwing simulation for 30 iterations and Table 9 has the average values for each of those iterations.
300
Ingalls
Table 9: 30 Iterations of Throwing Dice 100 Times

6.72 6.95 6.78 7.14 6.62 6.81
6.75 7.17 6.62 7.3 6.92 7.04
6.79 7.13 7.17 7.12 6.82 7.29
7.13 7.26 7.19 6.52 6.8 7.3
6.95 6.96 6.78 7.17 7.13 6.68
Average: 6.967
Standard Deviation: 0.22992
95% Confidence Interval: [6.8847,7.0493]
If we are simply trying to estimate the average of 100 rolls of the dice, we do not need to worry about
the standard deviation for each iteration. However, we do need to worry about the standard deviation of
the averages from each iteration. As is shown in Table 9, the average of the averages (so to speak) is
6.967. The standard deviation of those 30 averages is 0.22992. With that information, we can calculate a
confidence interval.
A confidence interval is a statistical measure where we want to bound some statistic. The level of
confidence (95% in our example) is the statistical probability that the statistic that we are considering lies
in the interval. So, the interpretation of the 95% confidence interval of [6.8847,7.0493] would be There
is a 95% chance that the true average of rolling the dice 100 times lies between 6.8847 and 7.0493.
Most statistical and simulation packages automatically calculate confidence intervals for you. Even Mi-
crosoft Excel has a function to calculate confidence intervals.
So in our example, we want to run 30 iterations so that we can have good confidence intervals for
each of our statistics. (Although we will not get into the topic in this paper, one should run 30 iterations
(or more) if you can in order to get good confidence intervals.) Table 10 shows the confidence intervals
for each of our statistics.
Table 10: Confidence Intervals after 30 Iterations

CI CI
Statistic Value Lower Upper
Revenue $1624.25 $1598.50 $1650.00
Lost Revenue $200.77 $178.71 $222.81
Lost Customers 17.07 15.34 18.79
MenuBoard Utilization 96.38% 95.56% 97.18%
Kitchen Utilization 70.19% 69.31% 71.06%
OrderWindow Utilization 98.53% 98.32% 98.75%
MenuBoard Waiting in Line 0.961 0.909 1.013
Kitchen Waiting in Line 0.163 0.152 0.174
OrderWindow Waiting in Line 1.241 1.200 1.282
Time In System 5.48 5.37 5.95
These statistics are very interesting on several fronts. First, could improve our revenue every day at
lunch 12.4% if we could capture the lost revenue. We need to find a way to lower the number of lost cus-
tomers from the 17.07 that we have seen in the current system. We also know that both the Menu Board
and Order Window are highly utilized, which means that lines are forming in front of those two resources.
We would also want to look at reducing the cars time in the system. The minimum average theoretical
time for a car would be 70 seconds at the Menu Board and 70 seconds at the Order Window, which is 140
301
Ingalls
seconds, or 2.33 minutes. So, about 3 minutes of the cars time is simply waiting. A key customer satis-
faction metric is to minimize the amount of time they have to wait in line.
6 FINDING WAYS TO IMPROVE A SYSTEM
What we have accomplished to this point is the analysis of a system that exists. We have learned that we
can improve revenue and that we have very busy resources. So let us run some alternative scenarios to
determine what (if any) improvements we should make.
6.1 Scenario 1: No Lines
Scenario 1 is based on the eliminating inventory thought. If we want to get the customer out of the sys-
tem quicker, we would simply need to eliminate some (or all) of the line. Well take drastic action and
eliminate all waiting in the system. If a car cannot pull right up to the Menu Board, then it will leave. If
the Order Window is not available after the order is placed, then the car waits at the Menu Board until the
Order Window is free. This should greatly lower the Time in System statistic. Some would say that this
would be the best way to run the drive-through line since the system is balanced. The arrival rate is 1
every 70 seconds, the Menu Board rate is 1 every 70 seconds, and the Order Window rate is 1 every 70
seconds. Some would say that this is a perfect production line.
In order to accomplish this in the model, we simply change 2 variables, MenuBoardLineSize and Or-
derPickupLineSize, and set them both to a value of 1. Then we run the model for 30 iterations and get the
statistics in Table 11.
Table 11: Scenario 1 Statistics

CI CI
Revenue $982.46 $939.89 $985.06
Lost Revenue $890.08 $860.72 $919.42
Well, the strategy worked the way we expected it to work. We reduced the Time in System statistic
by 38.3%. Now the cars are whipping through the drive-through. But there is a catch. Were losing 4.3
times the number of customers that we were losing before! Weve lowered our revenues by 40%! This
would hardly be considered a viable alternative!
What was it about eliminating the lines that caused such a problem? Without the lines, we caused ex-
cessive blocking in the system. Blocking occurs when something in the system will not allow a resource
to become available even though its service is finished. Throughout this scenario, the car at the
MenuBoard was finished ordering, but the car could not move forward because the Order Window was
still occupied. In this scenario (or in any balanced system with random variation and no queuing space),
each car at the MenuBoard has a 50% chance of being blocked by the OrderWindow. This can be re-
duced by adding queues between the resources.
302
Ingalls
In industrial problems, it is often difficult to determine where and why blocking occurs. That is why
there is so much emphasis on eliminating and/or controlling the bottleneck of the system. It is likely that
the bottleneck is in large part responsible for blocking in other parts of the system.
6.2 Scenario 2: Longer Lines at the Menu Board
Well, since Scenario 1 did not work very well, let us try another. Let us rearrange the parking lot so that
we can get room for 6 cars waiting in line for the Menu Board (including the car whose order is being
taken). We will keep the line for 3 cars for the Order Pickup Window. The results for Scenario 2 is
shown in Table 12.

CI CI
Revenue $1641.51 $1632.05 $1670.98
Lost Revenue $141.98 $120.72 $163.23
Well, we did add revenue ($17.26 per day) and decrease the number of lost customers, that is good,
but now the average customer waits an additional minute to get the order, which is bad. You can see why
this occurs. With the Menu Board line being longer, more cars can wait for the menu board and they do
not become lost customers in this scenario, where they would have become lost customers in the original
model. The price they pay for having room to get in line is that the line is longer, and so the time through
the system increases. This is another typical trade-off. Queues are good because they allow more
throughput in a system, but they add cycle time for the system.
6.3 Scenario 3: Improved Service Times
Let us try one more scenario. This scenario has found new technology that will cut the average service
time at the Menu Board and the Order Window by 20% from 70 seconds to 56 seconds. To implement
this technology, it would take $30,000 at each store and must be paid for by increased revenue at the
store. This is the only thing that changes from the original model. The results are shown in Table 13.
303
Ingalls

CI CI
Revenue $1773.66 $1747.74 $1799.60
Lost Revenue $39.63 $27.35 $51.90
Well, now we see some improvement! The Time in System has dropped from 5.48 minutes to 3.98
minutes. We have virtually eliminated any lost customers, and have increased revenue by $149.41 per
lunch shift. We would pay back the $30,000 for the implementation of the new technology in 200 days or
less than 7 months! It is well worth the investment and you are a hero for figuring it out!
Why did this scenario show such improvement? It was because our resources were not so highly
loaded. The Menu Board utilization dropped from 96% to 84%, and so the line for the Menu Board
would be much smaller. We would have seen a similar drop in the Order Window utilization, but now the
Order Window is being often being blocked by the Kitchen.
6.4 Generating Other Scenarios
This simulation and any other simulation can be used to evaluate many different scenarios if the person
creating the model allows some flexibility in the model structure. One also has to consider the amount of
time it takes to run a new scenario. In a large scale simulation, to run and evaluate a new scenario could
take several days.
7 WHAT HAVE WE LEARNED?
This exercise points out several things about simulation in general. First, simulation can mimic the dy-
namic behavior of a system. That is what it is built to do. Regardless of how complex a system may be,
it is likely that a simulation expert will be able to create a model that will evaluate it. However, the more
complex a system is, the longer it takes to model, run and evaluate. But do not be discouraged, there are
very good simulation people available to model large systems.
Second, you (or the person analyzing the system) must have a good understanding of simulation sta-
tistics. It is important during the creation of the model so that input distributions are used properly. It is
important during the analysis of the output statistics so that the output is not misinterpreted. Mistakes
with either the inputs or the outputs will cause the simulation analysis to be invalid.
Third, to analyze a system, simulation is used to evaluate different scenarios. It does not choose the
best scenario for you. This may seem to be a problem, but most managers have no shortage of scenarios
to evaluate. The trade off for this is that you can analyze the dynamics of the system and not just the av-
erage behavior.
Fourth, the scenarios that you do choose are generated by you and not the system. This is where fa-
miliarity with the system under study and a familiarity with system dynamics concepts are very valuable.
304
Ingalls
This is, of course, simply an introduction. Through this conference and interaction with simulation
professionals, you can get a deeper understanding of simulation and what it can do for you.
REFERENCES
Banks, J., J.S. Carson II, B.L. Nelson, and D.M. Nicol, 2009. Discrete Event System Simulation, 5th Ed.,
Upper Saddle River, New Jersey, Prentice-Hall.
Law, A.M., 2007. Simulation Modeling and Analysis, 4th Ed., New York, McGraw-Hill.
Kelton, W.D., R. Sadowski, D. Sadowski, 2001, Simulation with Arena, 2nd Edition, McGraw-Hill.
Ingalls, R.G, C. Kasales, 1999. CSCAT: Compaq Supply Chain Analysis Tool. Proceedings of the 1999
Winter Simulation Conference. ed. P.A. Farrington, H.B. Nembhard, D.T. Sturrock and G.W. Evans.
Piscataway, New Jersey: Institute of Electrical and Electronics Engineers.
Ingalls, R.G., 1998. The Value of Simulation in Modeling Supply Chains. Proceedings of the 1998 Win-
ter Simulation Conference. ed. D.J. Medeiros, E.F. Watson, J.S. Carson, and M.S. Manivannan. Pis-
cataway, New Jersey: Institute of Electrical and Electronics Engineers.
Ingalls, R.G., C. Eckersley, 1992. Simulation Issues in Electronics Manufacturing. Proceedings of the
1992 Winter Simulation Conference. ed. J.J. Swain, D. Goldsman, R.C. Crain and J.R. Wilson. Pisca-
taway, New Jersey: Institute of Electrical and Electronics Engineers.
Shannon, Robert E., 1975. Systems Simulation The Art and Science, Prentice-Hall.
AUTHOR BIOGRAPHY
RICKI G. INGALLS is Associate Professor and former Site Director of the Center for Engineering Lo-
gistics and Distribution (CELDi) in the School of Industrial Engineering and Management at Oklahoma
State University. He is also a Founding Principal with Diamond Head Associates, Inc. He joined OSU in
the Fall of 2000 after 16 years in industry with Compaq, SEMATECH, General Electric and Motorola.
He has a B.S. in Mathematics from East Texas Baptist College (1982), a M.S. in Industrial Engineering
from Texas A&M University (1984) and a Ph.D. in Management Science from the University of Texas at
Austin (1999). His research interests include the supply chain design issues and the development and ap-
plication of qualitative discrete-event simulation. He is a member of IIE. His email address is
<ricki.ingalls@okstate.edu>.
305

Introduction To Simulation

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Introduction To Simulation

Enviado por

Direitos autorais:

Formatos disponíveis

Proceedings of the 2013 Winter Simulation Conference

R. Pasupathy, S.-H. Kim, A. Tolk, R. Hill, and M. E. Kuhl, eds

Oklahoma State University

978-1-4799-2076-1/13/$31.00 2013 IEEE 291

Get in the Line for

When Arrive at the

Is there Room Wait until the Car

Get in Line for

Figure 1: Drive-Through Example Logic Flow

3 DISCRETE-EVENT SIMULATION STRUCTURE

3.2 Activities and Events

3.4 Global Variables

3.5 Random Number Generator

3.6 The Calendar

3.7 System State Variables

3.8 Statistics Collectors

4 A WALK THROUGH THESE CONCEPTS

4.1 The State of The System at Noon

Table 1: The Calendar at Noon

Entity Event Event Time

Table 2: Entity Attributes at Noon

Table 3: Statistics at Noon

4.2 The State of The System at 12:00:38 PM

Table 4: The Calendar at 12:00:38 PM

Table 5: Statistics at 12:00:38 PM

4.3 The State of The System at 12:00:46 PM

Lets take one more step in the simulation.

Table 6: Statistics at 12:00:46 PM

5 INTERPRETING OUTPUT STATISTICS

Table 7: Statistics at End of Simulation

Table 8: 100 Rolls of the Dice

Table 9: 30 Iterations of Throwing Dice 100 Times

Table 10: Confidence Intervals after 30 Iterations

6 FINDING WAYS TO IMPROVE A SYSTEM

6.1 Scenario 1: No Lines

Table 11: Scenario 1 Statistics

6.2 Scenario 2: Longer Lines at the Menu Board

Table 12: Scenario 2 Statistics

6.3 Scenario 3: Improved Service Times

Table 13: Scenario 3 Statistics

6.4 Generating Other Scenarios

7 WHAT HAVE WE LEARNED?

Você também pode gostar